Judge the residue sweeps on how much of the copy came back

@mjungnickel18 reports a third damage shape at the same broken residue (#25): after
an empty answer and a one-character answer, 400 tokens of fluent Danish that open with
a malformed <think> and invent a translation task, 2 of 1146 characters matching the
document. bugb_sweep filed that as DIVERGED -- the same class of miss as the
collapse-to-stop it had already been fixed for. The location is deterministic; the
damage is not, so no rule keyed on a signature can work.

bench/verbatim.py holds the rule both sweeps now use: a length floor, coverage -- the
fraction of the answer's 40-character windows that occur in the source -- against the
median of the other lengths in the same run, and a check for repetition the model added
beyond what the source itself repeats. All three carry weight. The live residue-4 break
returns ONE character, that character occurs in the document, and its coverage is
therefore 1.00; a long enough loop clears the coverage bar the same way.

It self-tests against all three shapes and against the healthy answers a signature rule
condemns: a one-character divergence mid-copy, a one-line preamble, a document that
repeats a line.

  venv/bin/python bench/verbatim.py

residue_sweep now prints coverage and a sha on every row rather than only the broken
ones -- "verbatim against the neighbours" cannot be applied to a log that prints only
outliers, and with it the neighbours of a break are visibly byte-identical to each
other. Its usage line names the file that exists.

Also the arithmetic it leads with. Five DISTINCT samples miss one broken residue in
128 C(127,5)/C(128,5) = 123/128 = 96% of the time, not 82%; 82% is the figure for six
broken residues. Corrected by @mjungnickel18, and it makes the point stronger: the
five residues this repo sampled had a 3.9% chance of finding residue 4.